Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory subsystem support
4 *
5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
6 * Dave Hansen <haveblue@us.ibm.com>
7 *
8 * This file provides the necessary infrastructure to represent
9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/topology.h>
17#include <linux/capability.h>
18#include <linux/device.h>
19#include <linux/memory.h>
20#include <linux/memory_hotplug.h>
21#include <linux/mm.h>
22#include <linux/stat.h>
23#include <linux/slab.h>
24#include <linux/xarray.h>
25
26#include <linux/atomic.h>
27#include <linux/uaccess.h>
28
29#define MEMORY_CLASS_NAME "memory"
30
31static const char *const online_type_to_str[] = {
32 [MMOP_OFFLINE] = "offline",
33 [MMOP_ONLINE] = "online",
34 [MMOP_ONLINE_KERNEL] = "online_kernel",
35 [MMOP_ONLINE_MOVABLE] = "online_movable",
36};
37
38int mhp_online_type_from_str(const char *str)
39{
40 int i;
41
42 for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
43 if (sysfs_streq(str, online_type_to_str[i]))
44 return i;
45 }
46 return -EINVAL;
47}
48
49#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
50
51static int sections_per_block;
52
53static inline unsigned long memory_block_id(unsigned long section_nr)
54{
55 return section_nr / sections_per_block;
56}
57
58static inline unsigned long pfn_to_block_id(unsigned long pfn)
59{
60 return memory_block_id(pfn_to_section_nr(pfn));
61}
62
63static inline unsigned long phys_to_block_id(unsigned long phys)
64{
65 return pfn_to_block_id(PFN_DOWN(phys));
66}
67
68static int memory_subsys_online(struct device *dev);
69static int memory_subsys_offline(struct device *dev);
70
71static struct bus_type memory_subsys = {
72 .name = MEMORY_CLASS_NAME,
73 .dev_name = MEMORY_CLASS_NAME,
74 .online = memory_subsys_online,
75 .offline = memory_subsys_offline,
76};
77
78/*
79 * Memory blocks are cached in a local radix tree to avoid
80 * a costly linear search for the corresponding device on
81 * the subsystem bus.
82 */
83static DEFINE_XARRAY(memory_blocks);
84
85/*
86 * Memory groups, indexed by memory group id (mgid).
87 */
88static DEFINE_XARRAY_FLAGS(memory_groups, XA_FLAGS_ALLOC);
89#define MEMORY_GROUP_MARK_DYNAMIC XA_MARK_1
90
91static BLOCKING_NOTIFIER_HEAD(memory_chain);
92
93int register_memory_notifier(struct notifier_block *nb)
94{
95 return blocking_notifier_chain_register(&memory_chain, nb);
96}
97EXPORT_SYMBOL(register_memory_notifier);
98
99void unregister_memory_notifier(struct notifier_block *nb)
100{
101 blocking_notifier_chain_unregister(&memory_chain, nb);
102}
103EXPORT_SYMBOL(unregister_memory_notifier);
104
105static void memory_block_release(struct device *dev)
106{
107 struct memory_block *mem = to_memory_block(dev);
108
109 kfree(mem);
110}
111
112unsigned long __weak memory_block_size_bytes(void)
113{
114 return MIN_MEMORY_BLOCK_SIZE;
115}
116EXPORT_SYMBOL_GPL(memory_block_size_bytes);
117
118/*
119 * Show the first physical section index (number) of this memory block.
120 */
121static ssize_t phys_index_show(struct device *dev,
122 struct device_attribute *attr, char *buf)
123{
124 struct memory_block *mem = to_memory_block(dev);
125 unsigned long phys_index;
126
127 phys_index = mem->start_section_nr / sections_per_block;
128
129 return sysfs_emit(buf, "%08lx\n", phys_index);
130}
131
132/*
133 * Legacy interface that we cannot remove. Always indicate "removable"
134 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
135 */
136static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
137 char *buf)
138{
139 return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
140}
141
142/*
143 * online, offline, going offline, etc.
144 */
145static ssize_t state_show(struct device *dev, struct device_attribute *attr,
146 char *buf)
147{
148 struct memory_block *mem = to_memory_block(dev);
149 const char *output;
150
151 /*
152 * We can probably put these states in a nice little array
153 * so that they're not open-coded
154 */
155 switch (mem->state) {
156 case MEM_ONLINE:
157 output = "online";
158 break;
159 case MEM_OFFLINE:
160 output = "offline";
161 break;
162 case MEM_GOING_OFFLINE:
163 output = "going-offline";
164 break;
165 default:
166 WARN_ON(1);
167 return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
168 }
169
170 return sysfs_emit(buf, "%s\n", output);
171}
172
173int memory_notify(unsigned long val, void *v)
174{
175 return blocking_notifier_call_chain(&memory_chain, val, v);
176}
177
178#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
179static unsigned long memblk_nr_poison(struct memory_block *mem);
180#else
181static inline unsigned long memblk_nr_poison(struct memory_block *mem)
182{
183 return 0;
184}
185#endif
186
187static int memory_block_online(struct memory_block *mem)
188{
189 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
190 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
191 unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
192 struct zone *zone;
193 int ret;
194
195 if (memblk_nr_poison(mem))
196 return -EHWPOISON;
197
198 zone = zone_for_pfn_range(mem->online_type, mem->nid, mem->group,
199 start_pfn, nr_pages);
200
201 /*
202 * Although vmemmap pages have a different lifecycle than the pages
203 * they describe (they remain until the memory is unplugged), doing
204 * their initialization and accounting at memory onlining/offlining
205 * stage helps to keep accounting easier to follow - e.g vmemmaps
206 * belong to the same zone as the memory they backed.
207 */
208 if (nr_vmemmap_pages) {
209 ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
210 if (ret)
211 return ret;
212 }
213
214 ret = online_pages(start_pfn + nr_vmemmap_pages,
215 nr_pages - nr_vmemmap_pages, zone, mem->group);
216 if (ret) {
217 if (nr_vmemmap_pages)
218 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
219 return ret;
220 }
221
222 /*
223 * Account once onlining succeeded. If the zone was unpopulated, it is
224 * now already properly populated.
225 */
226 if (nr_vmemmap_pages)
227 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
228 nr_vmemmap_pages);
229
230 mem->zone = zone;
231 return ret;
232}
233
234static int memory_block_offline(struct memory_block *mem)
235{
236 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
237 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
238 unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
239 int ret;
240
241 if (!mem->zone)
242 return -EINVAL;
243
244 /*
245 * Unaccount before offlining, such that unpopulated zone and kthreads
246 * can properly be torn down in offline_pages().
247 */
248 if (nr_vmemmap_pages)
249 adjust_present_page_count(pfn_to_page(start_pfn), mem->group,
250 -nr_vmemmap_pages);
251
252 ret = offline_pages(start_pfn + nr_vmemmap_pages,
253 nr_pages - nr_vmemmap_pages, mem->zone, mem->group);
254 if (ret) {
255 /* offline_pages() failed. Account back. */
256 if (nr_vmemmap_pages)
257 adjust_present_page_count(pfn_to_page(start_pfn),
258 mem->group, nr_vmemmap_pages);
259 return ret;
260 }
261
262 if (nr_vmemmap_pages)
263 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
264
265 mem->zone = NULL;
266 return ret;
267}
268
269/*
270 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
271 * OK to have direct references to sparsemem variables in here.
272 */
273static int
274memory_block_action(struct memory_block *mem, unsigned long action)
275{
276 int ret;
277
278 switch (action) {
279 case MEM_ONLINE:
280 ret = memory_block_online(mem);
281 break;
282 case MEM_OFFLINE:
283 ret = memory_block_offline(mem);
284 break;
285 default:
286 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
287 "%ld\n", __func__, mem->start_section_nr, action, action);
288 ret = -EINVAL;
289 }
290
291 return ret;
292}
293
294static int memory_block_change_state(struct memory_block *mem,
295 unsigned long to_state, unsigned long from_state_req)
296{
297 int ret = 0;
298
299 if (mem->state != from_state_req)
300 return -EINVAL;
301
302 if (to_state == MEM_OFFLINE)
303 mem->state = MEM_GOING_OFFLINE;
304
305 ret = memory_block_action(mem, to_state);
306 mem->state = ret ? from_state_req : to_state;
307
308 return ret;
309}
310
311/* The device lock serializes operations on memory_subsys_[online|offline] */
312static int memory_subsys_online(struct device *dev)
313{
314 struct memory_block *mem = to_memory_block(dev);
315 int ret;
316
317 if (mem->state == MEM_ONLINE)
318 return 0;
319
320 /*
321 * When called via device_online() without configuring the online_type,
322 * we want to default to MMOP_ONLINE.
323 */
324 if (mem->online_type == MMOP_OFFLINE)
325 mem->online_type = MMOP_ONLINE;
326
327 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
328 mem->online_type = MMOP_OFFLINE;
329
330 return ret;
331}
332
333static int memory_subsys_offline(struct device *dev)
334{
335 struct memory_block *mem = to_memory_block(dev);
336
337 if (mem->state == MEM_OFFLINE)
338 return 0;
339
340 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
341}
342
343static ssize_t state_store(struct device *dev, struct device_attribute *attr,
344 const char *buf, size_t count)
345{
346 const int online_type = mhp_online_type_from_str(buf);
347 struct memory_block *mem = to_memory_block(dev);
348 int ret;
349
350 if (online_type < 0)
351 return -EINVAL;
352
353 ret = lock_device_hotplug_sysfs();
354 if (ret)
355 return ret;
356
357 switch (online_type) {
358 case MMOP_ONLINE_KERNEL:
359 case MMOP_ONLINE_MOVABLE:
360 case MMOP_ONLINE:
361 /* mem->online_type is protected by device_hotplug_lock */
362 mem->online_type = online_type;
363 ret = device_online(&mem->dev);
364 break;
365 case MMOP_OFFLINE:
366 ret = device_offline(&mem->dev);
367 break;
368 default:
369 ret = -EINVAL; /* should never happen */
370 }
371
372 unlock_device_hotplug();
373
374 if (ret < 0)
375 return ret;
376 if (ret)
377 return -EINVAL;
378
379 return count;
380}
381
382/*
383 * Legacy interface that we cannot remove: s390x exposes the storage increment
384 * covered by a memory block, allowing for identifying which memory blocks
385 * comprise a storage increment. Since a memory block spans complete
386 * storage increments nowadays, this interface is basically unused. Other
387 * archs never exposed != 0.
388 */
389static ssize_t phys_device_show(struct device *dev,
390 struct device_attribute *attr, char *buf)
391{
392 struct memory_block *mem = to_memory_block(dev);
393 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
394
395 return sysfs_emit(buf, "%d\n",
396 arch_get_memory_phys_device(start_pfn));
397}
398
399#ifdef CONFIG_MEMORY_HOTREMOVE
400static int print_allowed_zone(char *buf, int len, int nid,
401 struct memory_group *group,
402 unsigned long start_pfn, unsigned long nr_pages,
403 int online_type, struct zone *default_zone)
404{
405 struct zone *zone;
406
407 zone = zone_for_pfn_range(online_type, nid, group, start_pfn, nr_pages);
408 if (zone == default_zone)
409 return 0;
410
411 return sysfs_emit_at(buf, len, " %s", zone->name);
412}
413
414static ssize_t valid_zones_show(struct device *dev,
415 struct device_attribute *attr, char *buf)
416{
417 struct memory_block *mem = to_memory_block(dev);
418 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
419 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
420 struct memory_group *group = mem->group;
421 struct zone *default_zone;
422 int nid = mem->nid;
423 int len = 0;
424
425 /*
426 * Check the existing zone. Make sure that we do that only on the
427 * online nodes otherwise the page_zone is not reliable
428 */
429 if (mem->state == MEM_ONLINE) {
430 /*
431 * If !mem->zone, the memory block spans multiple zones and
432 * cannot get offlined.
433 */
434 default_zone = mem->zone;
435 if (!default_zone)
436 return sysfs_emit(buf, "%s\n", "none");
437 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
438 goto out;
439 }
440
441 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, group,
442 start_pfn, nr_pages);
443
444 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
445 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
446 MMOP_ONLINE_KERNEL, default_zone);
447 len += print_allowed_zone(buf, len, nid, group, start_pfn, nr_pages,
448 MMOP_ONLINE_MOVABLE, default_zone);
449out:
450 len += sysfs_emit_at(buf, len, "\n");
451 return len;
452}
453static DEVICE_ATTR_RO(valid_zones);
454#endif
455
456static DEVICE_ATTR_RO(phys_index);
457static DEVICE_ATTR_RW(state);
458static DEVICE_ATTR_RO(phys_device);
459static DEVICE_ATTR_RO(removable);
460
461/*
462 * Show the memory block size (shared by all memory blocks).
463 */
464static ssize_t block_size_bytes_show(struct device *dev,
465 struct device_attribute *attr, char *buf)
466{
467 return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
468}
469
470static DEVICE_ATTR_RO(block_size_bytes);
471
472/*
473 * Memory auto online policy.
474 */
475
476static ssize_t auto_online_blocks_show(struct device *dev,
477 struct device_attribute *attr, char *buf)
478{
479 return sysfs_emit(buf, "%s\n",
480 online_type_to_str[mhp_default_online_type]);
481}
482
483static ssize_t auto_online_blocks_store(struct device *dev,
484 struct device_attribute *attr,
485 const char *buf, size_t count)
486{
487 const int online_type = mhp_online_type_from_str(buf);
488
489 if (online_type < 0)
490 return -EINVAL;
491
492 mhp_default_online_type = online_type;
493 return count;
494}
495
496static DEVICE_ATTR_RW(auto_online_blocks);
497
498/*
499 * Some architectures will have custom drivers to do this, and
500 * will not need to do it from userspace. The fake hot-add code
501 * as well as ppc64 will do all of their discovery in userspace
502 * and will require this interface.
503 */
504#ifdef CONFIG_ARCH_MEMORY_PROBE
505static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
506 const char *buf, size_t count)
507{
508 u64 phys_addr;
509 int nid, ret;
510 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
511
512 ret = kstrtoull(buf, 0, &phys_addr);
513 if (ret)
514 return ret;
515
516 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
517 return -EINVAL;
518
519 ret = lock_device_hotplug_sysfs();
520 if (ret)
521 return ret;
522
523 nid = memory_add_physaddr_to_nid(phys_addr);
524 ret = __add_memory(nid, phys_addr,
525 MIN_MEMORY_BLOCK_SIZE * sections_per_block,
526 MHP_NONE);
527
528 if (ret)
529 goto out;
530
531 ret = count;
532out:
533 unlock_device_hotplug();
534 return ret;
535}
536
537static DEVICE_ATTR_WO(probe);
538#endif
539
540#ifdef CONFIG_MEMORY_FAILURE
541/*
542 * Support for offlining pages of memory
543 */
544
545/* Soft offline a page */
546static ssize_t soft_offline_page_store(struct device *dev,
547 struct device_attribute *attr,
548 const char *buf, size_t count)
549{
550 int ret;
551 u64 pfn;
552 if (!capable(CAP_SYS_ADMIN))
553 return -EPERM;
554 if (kstrtoull(buf, 0, &pfn) < 0)
555 return -EINVAL;
556 pfn >>= PAGE_SHIFT;
557 ret = soft_offline_page(pfn, 0);
558 return ret == 0 ? count : ret;
559}
560
561/* Forcibly offline a page, including killing processes. */
562static ssize_t hard_offline_page_store(struct device *dev,
563 struct device_attribute *attr,
564 const char *buf, size_t count)
565{
566 int ret;
567 u64 pfn;
568 if (!capable(CAP_SYS_ADMIN))
569 return -EPERM;
570 if (kstrtoull(buf, 0, &pfn) < 0)
571 return -EINVAL;
572 pfn >>= PAGE_SHIFT;
573 ret = memory_failure(pfn, MF_SW_SIMULATED);
574 if (ret == -EOPNOTSUPP)
575 ret = 0;
576 return ret ? ret : count;
577}
578
579static DEVICE_ATTR_WO(soft_offline_page);
580static DEVICE_ATTR_WO(hard_offline_page);
581#endif
582
583/* See phys_device_show(). */
584int __weak arch_get_memory_phys_device(unsigned long start_pfn)
585{
586 return 0;
587}
588
589/*
590 * A reference for the returned memory block device is acquired.
591 *
592 * Called under device_hotplug_lock.
593 */
594static struct memory_block *find_memory_block_by_id(unsigned long block_id)
595{
596 struct memory_block *mem;
597
598 mem = xa_load(&memory_blocks, block_id);
599 if (mem)
600 get_device(&mem->dev);
601 return mem;
602}
603
604/*
605 * Called under device_hotplug_lock.
606 */
607struct memory_block *find_memory_block(unsigned long section_nr)
608{
609 unsigned long block_id = memory_block_id(section_nr);
610
611 return find_memory_block_by_id(block_id);
612}
613
614static struct attribute *memory_memblk_attrs[] = {
615 &dev_attr_phys_index.attr,
616 &dev_attr_state.attr,
617 &dev_attr_phys_device.attr,
618 &dev_attr_removable.attr,
619#ifdef CONFIG_MEMORY_HOTREMOVE
620 &dev_attr_valid_zones.attr,
621#endif
622 NULL
623};
624
625static const struct attribute_group memory_memblk_attr_group = {
626 .attrs = memory_memblk_attrs,
627};
628
629static const struct attribute_group *memory_memblk_attr_groups[] = {
630 &memory_memblk_attr_group,
631 NULL,
632};
633
634static int __add_memory_block(struct memory_block *memory)
635{
636 int ret;
637
638 memory->dev.bus = &memory_subsys;
639 memory->dev.id = memory->start_section_nr / sections_per_block;
640 memory->dev.release = memory_block_release;
641 memory->dev.groups = memory_memblk_attr_groups;
642 memory->dev.offline = memory->state == MEM_OFFLINE;
643
644 ret = device_register(&memory->dev);
645 if (ret) {
646 put_device(&memory->dev);
647 return ret;
648 }
649 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
650 GFP_KERNEL));
651 if (ret)
652 device_unregister(&memory->dev);
653
654 return ret;
655}
656
657static struct zone *early_node_zone_for_memory_block(struct memory_block *mem,
658 int nid)
659{
660 const unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
661 const unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
662 struct zone *zone, *matching_zone = NULL;
663 pg_data_t *pgdat = NODE_DATA(nid);
664 int i;
665
666 /*
667 * This logic only works for early memory, when the applicable zones
668 * already span the memory block. We don't expect overlapping zones on
669 * a single node for early memory. So if we're told that some PFNs
670 * of a node fall into this memory block, we can assume that all node
671 * zones that intersect with the memory block are actually applicable.
672 * No need to look at the memmap.
673 */
674 for (i = 0; i < MAX_NR_ZONES; i++) {
675 zone = pgdat->node_zones + i;
676 if (!populated_zone(zone))
677 continue;
678 if (!zone_intersects(zone, start_pfn, nr_pages))
679 continue;
680 if (!matching_zone) {
681 matching_zone = zone;
682 continue;
683 }
684 /* Spans multiple zones ... */
685 matching_zone = NULL;
686 break;
687 }
688 return matching_zone;
689}
690
691#ifdef CONFIG_NUMA
692/**
693 * memory_block_add_nid() - Indicate that system RAM falling into this memory
694 * block device (partially) belongs to the given node.
695 * @mem: The memory block device.
696 * @nid: The node id.
697 * @context: The memory initialization context.
698 *
699 * Indicate that system RAM falling into this memory block (partially) belongs
700 * to the given node. If the context indicates ("early") that we are adding the
701 * node during node device subsystem initialization, this will also properly
702 * set/adjust mem->zone based on the zone ranges of the given node.
703 */
704void memory_block_add_nid(struct memory_block *mem, int nid,
705 enum meminit_context context)
706{
707 if (context == MEMINIT_EARLY && mem->nid != nid) {
708 /*
709 * For early memory we have to determine the zone when setting
710 * the node id and handle multiple nodes spanning a single
711 * memory block by indicate via zone == NULL that we're not
712 * dealing with a single zone. So if we're setting the node id
713 * the first time, determine if there is a single zone. If we're
714 * setting the node id a second time to a different node,
715 * invalidate the single detected zone.
716 */
717 if (mem->nid == NUMA_NO_NODE)
718 mem->zone = early_node_zone_for_memory_block(mem, nid);
719 else
720 mem->zone = NULL;
721 }
722
723 /*
724 * If this memory block spans multiple nodes, we only indicate
725 * the last processed node. If we span multiple nodes (not applicable
726 * to hotplugged memory), zone == NULL will prohibit memory offlining
727 * and consequently unplug.
728 */
729 mem->nid = nid;
730}
731#endif
732
733static int add_memory_block(unsigned long block_id, unsigned long state,
734 unsigned long nr_vmemmap_pages,
735 struct memory_group *group)
736{
737 struct memory_block *mem;
738 int ret = 0;
739
740 mem = find_memory_block_by_id(block_id);
741 if (mem) {
742 put_device(&mem->dev);
743 return -EEXIST;
744 }
745 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
746 if (!mem)
747 return -ENOMEM;
748
749 mem->start_section_nr = block_id * sections_per_block;
750 mem->state = state;
751 mem->nid = NUMA_NO_NODE;
752 mem->nr_vmemmap_pages = nr_vmemmap_pages;
753 INIT_LIST_HEAD(&mem->group_next);
754
755#ifndef CONFIG_NUMA
756 if (state == MEM_ONLINE)
757 /*
758 * MEM_ONLINE at this point implies early memory. With NUMA,
759 * we'll determine the zone when setting the node id via
760 * memory_block_add_nid(). Memory hotplug updated the zone
761 * manually when memory onlining/offlining succeeds.
762 */
763 mem->zone = early_node_zone_for_memory_block(mem, NUMA_NO_NODE);
764#endif /* CONFIG_NUMA */
765
766 ret = __add_memory_block(mem);
767 if (ret)
768 return ret;
769
770 if (group) {
771 mem->group = group;
772 list_add(&mem->group_next, &group->memory_blocks);
773 }
774
775 return 0;
776}
777
778static int __init add_boot_memory_block(unsigned long base_section_nr)
779{
780 int section_count = 0;
781 unsigned long nr;
782
783 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
784 nr++)
785 if (present_section_nr(nr))
786 section_count++;
787
788 if (section_count == 0)
789 return 0;
790 return add_memory_block(memory_block_id(base_section_nr),
791 MEM_ONLINE, 0, NULL);
792}
793
794static int add_hotplug_memory_block(unsigned long block_id,
795 unsigned long nr_vmemmap_pages,
796 struct memory_group *group)
797{
798 return add_memory_block(block_id, MEM_OFFLINE, nr_vmemmap_pages, group);
799}
800
801static void remove_memory_block(struct memory_block *memory)
802{
803 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
804 return;
805
806 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
807
808 if (memory->group) {
809 list_del(&memory->group_next);
810 memory->group = NULL;
811 }
812
813 /* drop the ref. we got via find_memory_block() */
814 put_device(&memory->dev);
815 device_unregister(&memory->dev);
816}
817
818/*
819 * Create memory block devices for the given memory area. Start and size
820 * have to be aligned to memory block granularity. Memory block devices
821 * will be initialized as offline.
822 *
823 * Called under device_hotplug_lock.
824 */
825int create_memory_block_devices(unsigned long start, unsigned long size,
826 unsigned long vmemmap_pages,
827 struct memory_group *group)
828{
829 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
830 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
831 struct memory_block *mem;
832 unsigned long block_id;
833 int ret = 0;
834
835 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
836 !IS_ALIGNED(size, memory_block_size_bytes())))
837 return -EINVAL;
838
839 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
840 ret = add_hotplug_memory_block(block_id, vmemmap_pages, group);
841 if (ret)
842 break;
843 }
844 if (ret) {
845 end_block_id = block_id;
846 for (block_id = start_block_id; block_id != end_block_id;
847 block_id++) {
848 mem = find_memory_block_by_id(block_id);
849 if (WARN_ON_ONCE(!mem))
850 continue;
851 remove_memory_block(mem);
852 }
853 }
854 return ret;
855}
856
857/*
858 * Remove memory block devices for the given memory area. Start and size
859 * have to be aligned to memory block granularity. Memory block devices
860 * have to be offline.
861 *
862 * Called under device_hotplug_lock.
863 */
864void remove_memory_block_devices(unsigned long start, unsigned long size)
865{
866 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
867 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
868 struct memory_block *mem;
869 unsigned long block_id;
870
871 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
872 !IS_ALIGNED(size, memory_block_size_bytes())))
873 return;
874
875 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
876 mem = find_memory_block_by_id(block_id);
877 if (WARN_ON_ONCE(!mem))
878 continue;
879 num_poisoned_pages_sub(-1UL, memblk_nr_poison(mem));
880 unregister_memory_block_under_nodes(mem);
881 remove_memory_block(mem);
882 }
883}
884
885static struct attribute *memory_root_attrs[] = {
886#ifdef CONFIG_ARCH_MEMORY_PROBE
887 &dev_attr_probe.attr,
888#endif
889
890#ifdef CONFIG_MEMORY_FAILURE
891 &dev_attr_soft_offline_page.attr,
892 &dev_attr_hard_offline_page.attr,
893#endif
894
895 &dev_attr_block_size_bytes.attr,
896 &dev_attr_auto_online_blocks.attr,
897 NULL
898};
899
900static const struct attribute_group memory_root_attr_group = {
901 .attrs = memory_root_attrs,
902};
903
904static const struct attribute_group *memory_root_attr_groups[] = {
905 &memory_root_attr_group,
906 NULL,
907};
908
909/*
910 * Initialize the sysfs support for memory devices. At the time this function
911 * is called, we cannot have concurrent creation/deletion of memory block
912 * devices, the device_hotplug_lock is not needed.
913 */
914void __init memory_dev_init(void)
915{
916 int ret;
917 unsigned long block_sz, nr;
918
919 /* Validate the configured memory block size */
920 block_sz = memory_block_size_bytes();
921 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
922 panic("Memory block size not suitable: 0x%lx\n", block_sz);
923 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
924
925 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
926 if (ret)
927 panic("%s() failed to register subsystem: %d\n", __func__, ret);
928
929 /*
930 * Create entries for memory sections that were found
931 * during boot and have been initialized
932 */
933 for (nr = 0; nr <= __highest_present_section_nr;
934 nr += sections_per_block) {
935 ret = add_boot_memory_block(nr);
936 if (ret)
937 panic("%s() failed to add memory block: %d\n", __func__,
938 ret);
939 }
940}
941
942/**
943 * walk_memory_blocks - walk through all present memory blocks overlapped
944 * by the range [start, start + size)
945 *
946 * @start: start address of the memory range
947 * @size: size of the memory range
948 * @arg: argument passed to func
949 * @func: callback for each memory section walked
950 *
951 * This function walks through all present memory blocks overlapped by the
952 * range [start, start + size), calling func on each memory block.
953 *
954 * In case func() returns an error, walking is aborted and the error is
955 * returned.
956 *
957 * Called under device_hotplug_lock.
958 */
959int walk_memory_blocks(unsigned long start, unsigned long size,
960 void *arg, walk_memory_blocks_func_t func)
961{
962 const unsigned long start_block_id = phys_to_block_id(start);
963 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
964 struct memory_block *mem;
965 unsigned long block_id;
966 int ret = 0;
967
968 if (!size)
969 return 0;
970
971 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
972 mem = find_memory_block_by_id(block_id);
973 if (!mem)
974 continue;
975
976 ret = func(mem, arg);
977 put_device(&mem->dev);
978 if (ret)
979 break;
980 }
981 return ret;
982}
983
984struct for_each_memory_block_cb_data {
985 walk_memory_blocks_func_t func;
986 void *arg;
987};
988
989static int for_each_memory_block_cb(struct device *dev, void *data)
990{
991 struct memory_block *mem = to_memory_block(dev);
992 struct for_each_memory_block_cb_data *cb_data = data;
993
994 return cb_data->func(mem, cb_data->arg);
995}
996
997/**
998 * for_each_memory_block - walk through all present memory blocks
999 *
1000 * @arg: argument passed to func
1001 * @func: callback for each memory block walked
1002 *
1003 * This function walks through all present memory blocks, calling func on
1004 * each memory block.
1005 *
1006 * In case func() returns an error, walking is aborted and the error is
1007 * returned.
1008 */
1009int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
1010{
1011 struct for_each_memory_block_cb_data cb_data = {
1012 .func = func,
1013 .arg = arg,
1014 };
1015
1016 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
1017 for_each_memory_block_cb);
1018}
1019
1020/*
1021 * This is an internal helper to unify allocation and initialization of
1022 * memory groups. Note that the passed memory group will be copied to a
1023 * dynamically allocated memory group. After this call, the passed
1024 * memory group should no longer be used.
1025 */
1026static int memory_group_register(struct memory_group group)
1027{
1028 struct memory_group *new_group;
1029 uint32_t mgid;
1030 int ret;
1031
1032 if (!node_possible(group.nid))
1033 return -EINVAL;
1034
1035 new_group = kzalloc(sizeof(group), GFP_KERNEL);
1036 if (!new_group)
1037 return -ENOMEM;
1038 *new_group = group;
1039 INIT_LIST_HEAD(&new_group->memory_blocks);
1040
1041 ret = xa_alloc(&memory_groups, &mgid, new_group, xa_limit_31b,
1042 GFP_KERNEL);
1043 if (ret) {
1044 kfree(new_group);
1045 return ret;
1046 } else if (group.is_dynamic) {
1047 xa_set_mark(&memory_groups, mgid, MEMORY_GROUP_MARK_DYNAMIC);
1048 }
1049 return mgid;
1050}
1051
1052/**
1053 * memory_group_register_static() - Register a static memory group.
1054 * @nid: The node id.
1055 * @max_pages: The maximum number of pages we'll have in this static memory
1056 * group.
1057 *
1058 * Register a new static memory group and return the memory group id.
1059 * All memory in the group belongs to a single unit, such as a DIMM. All
1060 * memory belonging to a static memory group is added in one go to be removed
1061 * in one go -- it's static.
1062 *
1063 * Returns an error if out of memory, if the node id is invalid, if no new
1064 * memory groups can be registered, or if max_pages is invalid (0). Otherwise,
1065 * returns the new memory group id.
1066 */
1067int memory_group_register_static(int nid, unsigned long max_pages)
1068{
1069 struct memory_group group = {
1070 .nid = nid,
1071 .s = {
1072 .max_pages = max_pages,
1073 },
1074 };
1075
1076 if (!max_pages)
1077 return -EINVAL;
1078 return memory_group_register(group);
1079}
1080EXPORT_SYMBOL_GPL(memory_group_register_static);
1081
1082/**
1083 * memory_group_register_dynamic() - Register a dynamic memory group.
1084 * @nid: The node id.
1085 * @unit_pages: Unit in pages in which is memory added/removed in this dynamic
1086 * memory group.
1087 *
1088 * Register a new dynamic memory group and return the memory group id.
1089 * Memory within a dynamic memory group is added/removed dynamically
1090 * in unit_pages.
1091 *
1092 * Returns an error if out of memory, if the node id is invalid, if no new
1093 * memory groups can be registered, or if unit_pages is invalid (0, not a
1094 * power of two, smaller than a single memory block). Otherwise, returns the
1095 * new memory group id.
1096 */
1097int memory_group_register_dynamic(int nid, unsigned long unit_pages)
1098{
1099 struct memory_group group = {
1100 .nid = nid,
1101 .is_dynamic = true,
1102 .d = {
1103 .unit_pages = unit_pages,
1104 },
1105 };
1106
1107 if (!unit_pages || !is_power_of_2(unit_pages) ||
1108 unit_pages < PHYS_PFN(memory_block_size_bytes()))
1109 return -EINVAL;
1110 return memory_group_register(group);
1111}
1112EXPORT_SYMBOL_GPL(memory_group_register_dynamic);
1113
1114/**
1115 * memory_group_unregister() - Unregister a memory group.
1116 * @mgid: the memory group id
1117 *
1118 * Unregister a memory group. If any memory block still belongs to this
1119 * memory group, unregistering will fail.
1120 *
1121 * Returns -EINVAL if the memory group id is invalid, returns -EBUSY if some
1122 * memory blocks still belong to this memory group and returns 0 if
1123 * unregistering succeeded.
1124 */
1125int memory_group_unregister(int mgid)
1126{
1127 struct memory_group *group;
1128
1129 if (mgid < 0)
1130 return -EINVAL;
1131
1132 group = xa_load(&memory_groups, mgid);
1133 if (!group)
1134 return -EINVAL;
1135 if (!list_empty(&group->memory_blocks))
1136 return -EBUSY;
1137 xa_erase(&memory_groups, mgid);
1138 kfree(group);
1139 return 0;
1140}
1141EXPORT_SYMBOL_GPL(memory_group_unregister);
1142
1143/*
1144 * This is an internal helper only to be used in core memory hotplug code to
1145 * lookup a memory group. We don't care about locking, as we don't expect a
1146 * memory group to get unregistered while adding memory to it -- because
1147 * the group and the memory is managed by the same driver.
1148 */
1149struct memory_group *memory_group_find_by_id(int mgid)
1150{
1151 return xa_load(&memory_groups, mgid);
1152}
1153
1154/*
1155 * This is an internal helper only to be used in core memory hotplug code to
1156 * walk all dynamic memory groups excluding a given memory group, either
1157 * belonging to a specific node, or belonging to any node.
1158 */
1159int walk_dynamic_memory_groups(int nid, walk_memory_groups_func_t func,
1160 struct memory_group *excluded, void *arg)
1161{
1162 struct memory_group *group;
1163 unsigned long index;
1164 int ret = 0;
1165
1166 xa_for_each_marked(&memory_groups, index, group,
1167 MEMORY_GROUP_MARK_DYNAMIC) {
1168 if (group == excluded)
1169 continue;
1170#ifdef CONFIG_NUMA
1171 if (nid != NUMA_NO_NODE && group->nid != nid)
1172 continue;
1173#endif /* CONFIG_NUMA */
1174 ret = func(group, arg);
1175 if (ret)
1176 break;
1177 }
1178 return ret;
1179}
1180
1181#if defined(CONFIG_MEMORY_FAILURE) && defined(CONFIG_MEMORY_HOTPLUG)
1182void memblk_nr_poison_inc(unsigned long pfn)
1183{
1184 const unsigned long block_id = pfn_to_block_id(pfn);
1185 struct memory_block *mem = find_memory_block_by_id(block_id);
1186
1187 if (mem)
1188 atomic_long_inc(&mem->nr_hwpoison);
1189}
1190
1191void memblk_nr_poison_sub(unsigned long pfn, long i)
1192{
1193 const unsigned long block_id = pfn_to_block_id(pfn);
1194 struct memory_block *mem = find_memory_block_by_id(block_id);
1195
1196 if (mem)
1197 atomic_long_sub(i, &mem->nr_hwpoison);
1198}
1199
1200static unsigned long memblk_nr_poison(struct memory_block *mem)
1201{
1202 return atomic_long_read(&mem->nr_hwpoison);
1203}
1204#endif
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Memory subsystem support
4 *
5 * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
6 * Dave Hansen <haveblue@us.ibm.com>
7 *
8 * This file provides the necessary infrastructure to represent
9 * a SPARSEMEM-memory-model system's physical memory in /sysfs.
10 * All arch-independent code that assumes MEMORY_HOTPLUG requires
11 * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
12 */
13
14#include <linux/module.h>
15#include <linux/init.h>
16#include <linux/topology.h>
17#include <linux/capability.h>
18#include <linux/device.h>
19#include <linux/memory.h>
20#include <linux/memory_hotplug.h>
21#include <linux/mm.h>
22#include <linux/stat.h>
23#include <linux/slab.h>
24#include <linux/xarray.h>
25
26#include <linux/atomic.h>
27#include <linux/uaccess.h>
28
29#define MEMORY_CLASS_NAME "memory"
30
31static const char *const online_type_to_str[] = {
32 [MMOP_OFFLINE] = "offline",
33 [MMOP_ONLINE] = "online",
34 [MMOP_ONLINE_KERNEL] = "online_kernel",
35 [MMOP_ONLINE_MOVABLE] = "online_movable",
36};
37
38int mhp_online_type_from_str(const char *str)
39{
40 int i;
41
42 for (i = 0; i < ARRAY_SIZE(online_type_to_str); i++) {
43 if (sysfs_streq(str, online_type_to_str[i]))
44 return i;
45 }
46 return -EINVAL;
47}
48
49#define to_memory_block(dev) container_of(dev, struct memory_block, dev)
50
51static int sections_per_block;
52
53static inline unsigned long memory_block_id(unsigned long section_nr)
54{
55 return section_nr / sections_per_block;
56}
57
58static inline unsigned long pfn_to_block_id(unsigned long pfn)
59{
60 return memory_block_id(pfn_to_section_nr(pfn));
61}
62
63static inline unsigned long phys_to_block_id(unsigned long phys)
64{
65 return pfn_to_block_id(PFN_DOWN(phys));
66}
67
68static int memory_subsys_online(struct device *dev);
69static int memory_subsys_offline(struct device *dev);
70
71static struct bus_type memory_subsys = {
72 .name = MEMORY_CLASS_NAME,
73 .dev_name = MEMORY_CLASS_NAME,
74 .online = memory_subsys_online,
75 .offline = memory_subsys_offline,
76};
77
78/*
79 * Memory blocks are cached in a local radix tree to avoid
80 * a costly linear search for the corresponding device on
81 * the subsystem bus.
82 */
83static DEFINE_XARRAY(memory_blocks);
84
85static BLOCKING_NOTIFIER_HEAD(memory_chain);
86
87int register_memory_notifier(struct notifier_block *nb)
88{
89 return blocking_notifier_chain_register(&memory_chain, nb);
90}
91EXPORT_SYMBOL(register_memory_notifier);
92
93void unregister_memory_notifier(struct notifier_block *nb)
94{
95 blocking_notifier_chain_unregister(&memory_chain, nb);
96}
97EXPORT_SYMBOL(unregister_memory_notifier);
98
99static void memory_block_release(struct device *dev)
100{
101 struct memory_block *mem = to_memory_block(dev);
102
103 kfree(mem);
104}
105
106unsigned long __weak memory_block_size_bytes(void)
107{
108 return MIN_MEMORY_BLOCK_SIZE;
109}
110EXPORT_SYMBOL_GPL(memory_block_size_bytes);
111
112/*
113 * Show the first physical section index (number) of this memory block.
114 */
115static ssize_t phys_index_show(struct device *dev,
116 struct device_attribute *attr, char *buf)
117{
118 struct memory_block *mem = to_memory_block(dev);
119 unsigned long phys_index;
120
121 phys_index = mem->start_section_nr / sections_per_block;
122
123 return sysfs_emit(buf, "%08lx\n", phys_index);
124}
125
126/*
127 * Legacy interface that we cannot remove. Always indicate "removable"
128 * with CONFIG_MEMORY_HOTREMOVE - bad heuristic.
129 */
130static ssize_t removable_show(struct device *dev, struct device_attribute *attr,
131 char *buf)
132{
133 return sysfs_emit(buf, "%d\n", (int)IS_ENABLED(CONFIG_MEMORY_HOTREMOVE));
134}
135
136/*
137 * online, offline, going offline, etc.
138 */
139static ssize_t state_show(struct device *dev, struct device_attribute *attr,
140 char *buf)
141{
142 struct memory_block *mem = to_memory_block(dev);
143 const char *output;
144
145 /*
146 * We can probably put these states in a nice little array
147 * so that they're not open-coded
148 */
149 switch (mem->state) {
150 case MEM_ONLINE:
151 output = "online";
152 break;
153 case MEM_OFFLINE:
154 output = "offline";
155 break;
156 case MEM_GOING_OFFLINE:
157 output = "going-offline";
158 break;
159 default:
160 WARN_ON(1);
161 return sysfs_emit(buf, "ERROR-UNKNOWN-%ld\n", mem->state);
162 }
163
164 return sysfs_emit(buf, "%s\n", output);
165}
166
167int memory_notify(unsigned long val, void *v)
168{
169 return blocking_notifier_call_chain(&memory_chain, val, v);
170}
171
172static int memory_block_online(struct memory_block *mem)
173{
174 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
175 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
176 unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
177 struct zone *zone;
178 int ret;
179
180 zone = zone_for_pfn_range(mem->online_type, mem->nid, start_pfn, nr_pages);
181
182 /*
183 * Although vmemmap pages have a different lifecycle than the pages
184 * they describe (they remain until the memory is unplugged), doing
185 * their initialization and accounting at memory onlining/offlining
186 * stage helps to keep accounting easier to follow - e.g vmemmaps
187 * belong to the same zone as the memory they backed.
188 */
189 if (nr_vmemmap_pages) {
190 ret = mhp_init_memmap_on_memory(start_pfn, nr_vmemmap_pages, zone);
191 if (ret)
192 return ret;
193 }
194
195 ret = online_pages(start_pfn + nr_vmemmap_pages,
196 nr_pages - nr_vmemmap_pages, zone);
197 if (ret) {
198 if (nr_vmemmap_pages)
199 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
200 return ret;
201 }
202
203 /*
204 * Account once onlining succeeded. If the zone was unpopulated, it is
205 * now already properly populated.
206 */
207 if (nr_vmemmap_pages)
208 adjust_present_page_count(zone, nr_vmemmap_pages);
209
210 return ret;
211}
212
213static int memory_block_offline(struct memory_block *mem)
214{
215 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
216 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
217 unsigned long nr_vmemmap_pages = mem->nr_vmemmap_pages;
218 struct zone *zone;
219 int ret;
220
221 /*
222 * Unaccount before offlining, such that unpopulated zone and kthreads
223 * can properly be torn down in offline_pages().
224 */
225 if (nr_vmemmap_pages) {
226 zone = page_zone(pfn_to_page(start_pfn));
227 adjust_present_page_count(zone, -nr_vmemmap_pages);
228 }
229
230 ret = offline_pages(start_pfn + nr_vmemmap_pages,
231 nr_pages - nr_vmemmap_pages);
232 if (ret) {
233 /* offline_pages() failed. Account back. */
234 if (nr_vmemmap_pages)
235 adjust_present_page_count(zone, nr_vmemmap_pages);
236 return ret;
237 }
238
239 if (nr_vmemmap_pages)
240 mhp_deinit_memmap_on_memory(start_pfn, nr_vmemmap_pages);
241
242 return ret;
243}
244
245/*
246 * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
247 * OK to have direct references to sparsemem variables in here.
248 */
249static int
250memory_block_action(struct memory_block *mem, unsigned long action)
251{
252 int ret;
253
254 switch (action) {
255 case MEM_ONLINE:
256 ret = memory_block_online(mem);
257 break;
258 case MEM_OFFLINE:
259 ret = memory_block_offline(mem);
260 break;
261 default:
262 WARN(1, KERN_WARNING "%s(%ld, %ld) unknown action: "
263 "%ld\n", __func__, mem->start_section_nr, action, action);
264 ret = -EINVAL;
265 }
266
267 return ret;
268}
269
270static int memory_block_change_state(struct memory_block *mem,
271 unsigned long to_state, unsigned long from_state_req)
272{
273 int ret = 0;
274
275 if (mem->state != from_state_req)
276 return -EINVAL;
277
278 if (to_state == MEM_OFFLINE)
279 mem->state = MEM_GOING_OFFLINE;
280
281 ret = memory_block_action(mem, to_state);
282 mem->state = ret ? from_state_req : to_state;
283
284 return ret;
285}
286
287/* The device lock serializes operations on memory_subsys_[online|offline] */
288static int memory_subsys_online(struct device *dev)
289{
290 struct memory_block *mem = to_memory_block(dev);
291 int ret;
292
293 if (mem->state == MEM_ONLINE)
294 return 0;
295
296 /*
297 * When called via device_online() without configuring the online_type,
298 * we want to default to MMOP_ONLINE.
299 */
300 if (mem->online_type == MMOP_OFFLINE)
301 mem->online_type = MMOP_ONLINE;
302
303 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
304 mem->online_type = MMOP_OFFLINE;
305
306 return ret;
307}
308
309static int memory_subsys_offline(struct device *dev)
310{
311 struct memory_block *mem = to_memory_block(dev);
312
313 if (mem->state == MEM_OFFLINE)
314 return 0;
315
316 return memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
317}
318
319static ssize_t state_store(struct device *dev, struct device_attribute *attr,
320 const char *buf, size_t count)
321{
322 const int online_type = mhp_online_type_from_str(buf);
323 struct memory_block *mem = to_memory_block(dev);
324 int ret;
325
326 if (online_type < 0)
327 return -EINVAL;
328
329 ret = lock_device_hotplug_sysfs();
330 if (ret)
331 return ret;
332
333 switch (online_type) {
334 case MMOP_ONLINE_KERNEL:
335 case MMOP_ONLINE_MOVABLE:
336 case MMOP_ONLINE:
337 /* mem->online_type is protected by device_hotplug_lock */
338 mem->online_type = online_type;
339 ret = device_online(&mem->dev);
340 break;
341 case MMOP_OFFLINE:
342 ret = device_offline(&mem->dev);
343 break;
344 default:
345 ret = -EINVAL; /* should never happen */
346 }
347
348 unlock_device_hotplug();
349
350 if (ret < 0)
351 return ret;
352 if (ret)
353 return -EINVAL;
354
355 return count;
356}
357
358/*
359 * Legacy interface that we cannot remove: s390x exposes the storage increment
360 * covered by a memory block, allowing for identifying which memory blocks
361 * comprise a storage increment. Since a memory block spans complete
362 * storage increments nowadays, this interface is basically unused. Other
363 * archs never exposed != 0.
364 */
365static ssize_t phys_device_show(struct device *dev,
366 struct device_attribute *attr, char *buf)
367{
368 struct memory_block *mem = to_memory_block(dev);
369 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
370
371 return sysfs_emit(buf, "%d\n",
372 arch_get_memory_phys_device(start_pfn));
373}
374
375#ifdef CONFIG_MEMORY_HOTREMOVE
376static int print_allowed_zone(char *buf, int len, int nid,
377 unsigned long start_pfn, unsigned long nr_pages,
378 int online_type, struct zone *default_zone)
379{
380 struct zone *zone;
381
382 zone = zone_for_pfn_range(online_type, nid, start_pfn, nr_pages);
383 if (zone == default_zone)
384 return 0;
385
386 return sysfs_emit_at(buf, len, " %s", zone->name);
387}
388
389static ssize_t valid_zones_show(struct device *dev,
390 struct device_attribute *attr, char *buf)
391{
392 struct memory_block *mem = to_memory_block(dev);
393 unsigned long start_pfn = section_nr_to_pfn(mem->start_section_nr);
394 unsigned long nr_pages = PAGES_PER_SECTION * sections_per_block;
395 struct zone *default_zone;
396 int len = 0;
397 int nid;
398
399 /*
400 * Check the existing zone. Make sure that we do that only on the
401 * online nodes otherwise the page_zone is not reliable
402 */
403 if (mem->state == MEM_ONLINE) {
404 /*
405 * The block contains more than one zone can not be offlined.
406 * This can happen e.g. for ZONE_DMA and ZONE_DMA32
407 */
408 default_zone = test_pages_in_a_zone(start_pfn,
409 start_pfn + nr_pages);
410 if (!default_zone)
411 return sysfs_emit(buf, "%s\n", "none");
412 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
413 goto out;
414 }
415
416 nid = mem->nid;
417 default_zone = zone_for_pfn_range(MMOP_ONLINE, nid, start_pfn,
418 nr_pages);
419
420 len += sysfs_emit_at(buf, len, "%s", default_zone->name);
421 len += print_allowed_zone(buf, len, nid, start_pfn, nr_pages,
422 MMOP_ONLINE_KERNEL, default_zone);
423 len += print_allowed_zone(buf, len, nid, start_pfn, nr_pages,
424 MMOP_ONLINE_MOVABLE, default_zone);
425out:
426 len += sysfs_emit_at(buf, len, "\n");
427 return len;
428}
429static DEVICE_ATTR_RO(valid_zones);
430#endif
431
432static DEVICE_ATTR_RO(phys_index);
433static DEVICE_ATTR_RW(state);
434static DEVICE_ATTR_RO(phys_device);
435static DEVICE_ATTR_RO(removable);
436
437/*
438 * Show the memory block size (shared by all memory blocks).
439 */
440static ssize_t block_size_bytes_show(struct device *dev,
441 struct device_attribute *attr, char *buf)
442{
443 return sysfs_emit(buf, "%lx\n", memory_block_size_bytes());
444}
445
446static DEVICE_ATTR_RO(block_size_bytes);
447
448/*
449 * Memory auto online policy.
450 */
451
452static ssize_t auto_online_blocks_show(struct device *dev,
453 struct device_attribute *attr, char *buf)
454{
455 return sysfs_emit(buf, "%s\n",
456 online_type_to_str[mhp_default_online_type]);
457}
458
459static ssize_t auto_online_blocks_store(struct device *dev,
460 struct device_attribute *attr,
461 const char *buf, size_t count)
462{
463 const int online_type = mhp_online_type_from_str(buf);
464
465 if (online_type < 0)
466 return -EINVAL;
467
468 mhp_default_online_type = online_type;
469 return count;
470}
471
472static DEVICE_ATTR_RW(auto_online_blocks);
473
474/*
475 * Some architectures will have custom drivers to do this, and
476 * will not need to do it from userspace. The fake hot-add code
477 * as well as ppc64 will do all of their discovery in userspace
478 * and will require this interface.
479 */
480#ifdef CONFIG_ARCH_MEMORY_PROBE
481static ssize_t probe_store(struct device *dev, struct device_attribute *attr,
482 const char *buf, size_t count)
483{
484 u64 phys_addr;
485 int nid, ret;
486 unsigned long pages_per_block = PAGES_PER_SECTION * sections_per_block;
487
488 ret = kstrtoull(buf, 0, &phys_addr);
489 if (ret)
490 return ret;
491
492 if (phys_addr & ((pages_per_block << PAGE_SHIFT) - 1))
493 return -EINVAL;
494
495 ret = lock_device_hotplug_sysfs();
496 if (ret)
497 return ret;
498
499 nid = memory_add_physaddr_to_nid(phys_addr);
500 ret = __add_memory(nid, phys_addr,
501 MIN_MEMORY_BLOCK_SIZE * sections_per_block,
502 MHP_NONE);
503
504 if (ret)
505 goto out;
506
507 ret = count;
508out:
509 unlock_device_hotplug();
510 return ret;
511}
512
513static DEVICE_ATTR_WO(probe);
514#endif
515
516#ifdef CONFIG_MEMORY_FAILURE
517/*
518 * Support for offlining pages of memory
519 */
520
521/* Soft offline a page */
522static ssize_t soft_offline_page_store(struct device *dev,
523 struct device_attribute *attr,
524 const char *buf, size_t count)
525{
526 int ret;
527 u64 pfn;
528 if (!capable(CAP_SYS_ADMIN))
529 return -EPERM;
530 if (kstrtoull(buf, 0, &pfn) < 0)
531 return -EINVAL;
532 pfn >>= PAGE_SHIFT;
533 ret = soft_offline_page(pfn, 0);
534 return ret == 0 ? count : ret;
535}
536
537/* Forcibly offline a page, including killing processes. */
538static ssize_t hard_offline_page_store(struct device *dev,
539 struct device_attribute *attr,
540 const char *buf, size_t count)
541{
542 int ret;
543 u64 pfn;
544 if (!capable(CAP_SYS_ADMIN))
545 return -EPERM;
546 if (kstrtoull(buf, 0, &pfn) < 0)
547 return -EINVAL;
548 pfn >>= PAGE_SHIFT;
549 ret = memory_failure(pfn, 0);
550 return ret ? ret : count;
551}
552
553static DEVICE_ATTR_WO(soft_offline_page);
554static DEVICE_ATTR_WO(hard_offline_page);
555#endif
556
557/* See phys_device_show(). */
558int __weak arch_get_memory_phys_device(unsigned long start_pfn)
559{
560 return 0;
561}
562
563/*
564 * A reference for the returned memory block device is acquired.
565 *
566 * Called under device_hotplug_lock.
567 */
568static struct memory_block *find_memory_block_by_id(unsigned long block_id)
569{
570 struct memory_block *mem;
571
572 mem = xa_load(&memory_blocks, block_id);
573 if (mem)
574 get_device(&mem->dev);
575 return mem;
576}
577
578/*
579 * Called under device_hotplug_lock.
580 */
581struct memory_block *find_memory_block(struct mem_section *section)
582{
583 unsigned long block_id = memory_block_id(__section_nr(section));
584
585 return find_memory_block_by_id(block_id);
586}
587
588static struct attribute *memory_memblk_attrs[] = {
589 &dev_attr_phys_index.attr,
590 &dev_attr_state.attr,
591 &dev_attr_phys_device.attr,
592 &dev_attr_removable.attr,
593#ifdef CONFIG_MEMORY_HOTREMOVE
594 &dev_attr_valid_zones.attr,
595#endif
596 NULL
597};
598
599static const struct attribute_group memory_memblk_attr_group = {
600 .attrs = memory_memblk_attrs,
601};
602
603static const struct attribute_group *memory_memblk_attr_groups[] = {
604 &memory_memblk_attr_group,
605 NULL,
606};
607
608/*
609 * register_memory - Setup a sysfs device for a memory block
610 */
611static
612int register_memory(struct memory_block *memory)
613{
614 int ret;
615
616 memory->dev.bus = &memory_subsys;
617 memory->dev.id = memory->start_section_nr / sections_per_block;
618 memory->dev.release = memory_block_release;
619 memory->dev.groups = memory_memblk_attr_groups;
620 memory->dev.offline = memory->state == MEM_OFFLINE;
621
622 ret = device_register(&memory->dev);
623 if (ret) {
624 put_device(&memory->dev);
625 return ret;
626 }
627 ret = xa_err(xa_store(&memory_blocks, memory->dev.id, memory,
628 GFP_KERNEL));
629 if (ret) {
630 put_device(&memory->dev);
631 device_unregister(&memory->dev);
632 }
633 return ret;
634}
635
636static int init_memory_block(unsigned long block_id, unsigned long state,
637 unsigned long nr_vmemmap_pages)
638{
639 struct memory_block *mem;
640 int ret = 0;
641
642 mem = find_memory_block_by_id(block_id);
643 if (mem) {
644 put_device(&mem->dev);
645 return -EEXIST;
646 }
647 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
648 if (!mem)
649 return -ENOMEM;
650
651 mem->start_section_nr = block_id * sections_per_block;
652 mem->state = state;
653 mem->nid = NUMA_NO_NODE;
654 mem->nr_vmemmap_pages = nr_vmemmap_pages;
655
656 ret = register_memory(mem);
657
658 return ret;
659}
660
661static int add_memory_block(unsigned long base_section_nr)
662{
663 int section_count = 0;
664 unsigned long nr;
665
666 for (nr = base_section_nr; nr < base_section_nr + sections_per_block;
667 nr++)
668 if (present_section_nr(nr))
669 section_count++;
670
671 if (section_count == 0)
672 return 0;
673 return init_memory_block(memory_block_id(base_section_nr),
674 MEM_ONLINE, 0);
675}
676
677static void unregister_memory(struct memory_block *memory)
678{
679 if (WARN_ON_ONCE(memory->dev.bus != &memory_subsys))
680 return;
681
682 WARN_ON(xa_erase(&memory_blocks, memory->dev.id) == NULL);
683
684 /* drop the ref. we got via find_memory_block() */
685 put_device(&memory->dev);
686 device_unregister(&memory->dev);
687}
688
689/*
690 * Create memory block devices for the given memory area. Start and size
691 * have to be aligned to memory block granularity. Memory block devices
692 * will be initialized as offline.
693 *
694 * Called under device_hotplug_lock.
695 */
696int create_memory_block_devices(unsigned long start, unsigned long size,
697 unsigned long vmemmap_pages)
698{
699 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
700 unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
701 struct memory_block *mem;
702 unsigned long block_id;
703 int ret = 0;
704
705 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
706 !IS_ALIGNED(size, memory_block_size_bytes())))
707 return -EINVAL;
708
709 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
710 ret = init_memory_block(block_id, MEM_OFFLINE, vmemmap_pages);
711 if (ret)
712 break;
713 }
714 if (ret) {
715 end_block_id = block_id;
716 for (block_id = start_block_id; block_id != end_block_id;
717 block_id++) {
718 mem = find_memory_block_by_id(block_id);
719 if (WARN_ON_ONCE(!mem))
720 continue;
721 unregister_memory(mem);
722 }
723 }
724 return ret;
725}
726
727/*
728 * Remove memory block devices for the given memory area. Start and size
729 * have to be aligned to memory block granularity. Memory block devices
730 * have to be offline.
731 *
732 * Called under device_hotplug_lock.
733 */
734void remove_memory_block_devices(unsigned long start, unsigned long size)
735{
736 const unsigned long start_block_id = pfn_to_block_id(PFN_DOWN(start));
737 const unsigned long end_block_id = pfn_to_block_id(PFN_DOWN(start + size));
738 struct memory_block *mem;
739 unsigned long block_id;
740
741 if (WARN_ON_ONCE(!IS_ALIGNED(start, memory_block_size_bytes()) ||
742 !IS_ALIGNED(size, memory_block_size_bytes())))
743 return;
744
745 for (block_id = start_block_id; block_id != end_block_id; block_id++) {
746 mem = find_memory_block_by_id(block_id);
747 if (WARN_ON_ONCE(!mem))
748 continue;
749 unregister_memory_block_under_nodes(mem);
750 unregister_memory(mem);
751 }
752}
753
754/* return true if the memory block is offlined, otherwise, return false */
755bool is_memblock_offlined(struct memory_block *mem)
756{
757 return mem->state == MEM_OFFLINE;
758}
759
760static struct attribute *memory_root_attrs[] = {
761#ifdef CONFIG_ARCH_MEMORY_PROBE
762 &dev_attr_probe.attr,
763#endif
764
765#ifdef CONFIG_MEMORY_FAILURE
766 &dev_attr_soft_offline_page.attr,
767 &dev_attr_hard_offline_page.attr,
768#endif
769
770 &dev_attr_block_size_bytes.attr,
771 &dev_attr_auto_online_blocks.attr,
772 NULL
773};
774
775static const struct attribute_group memory_root_attr_group = {
776 .attrs = memory_root_attrs,
777};
778
779static const struct attribute_group *memory_root_attr_groups[] = {
780 &memory_root_attr_group,
781 NULL,
782};
783
784/*
785 * Initialize the sysfs support for memory devices. At the time this function
786 * is called, we cannot have concurrent creation/deletion of memory block
787 * devices, the device_hotplug_lock is not needed.
788 */
789void __init memory_dev_init(void)
790{
791 int ret;
792 unsigned long block_sz, nr;
793
794 /* Validate the configured memory block size */
795 block_sz = memory_block_size_bytes();
796 if (!is_power_of_2(block_sz) || block_sz < MIN_MEMORY_BLOCK_SIZE)
797 panic("Memory block size not suitable: 0x%lx\n", block_sz);
798 sections_per_block = block_sz / MIN_MEMORY_BLOCK_SIZE;
799
800 ret = subsys_system_register(&memory_subsys, memory_root_attr_groups);
801 if (ret)
802 panic("%s() failed to register subsystem: %d\n", __func__, ret);
803
804 /*
805 * Create entries for memory sections that were found
806 * during boot and have been initialized
807 */
808 for (nr = 0; nr <= __highest_present_section_nr;
809 nr += sections_per_block) {
810 ret = add_memory_block(nr);
811 if (ret)
812 panic("%s() failed to add memory block: %d\n", __func__,
813 ret);
814 }
815}
816
817/**
818 * walk_memory_blocks - walk through all present memory blocks overlapped
819 * by the range [start, start + size)
820 *
821 * @start: start address of the memory range
822 * @size: size of the memory range
823 * @arg: argument passed to func
824 * @func: callback for each memory section walked
825 *
826 * This function walks through all present memory blocks overlapped by the
827 * range [start, start + size), calling func on each memory block.
828 *
829 * In case func() returns an error, walking is aborted and the error is
830 * returned.
831 *
832 * Called under device_hotplug_lock.
833 */
834int walk_memory_blocks(unsigned long start, unsigned long size,
835 void *arg, walk_memory_blocks_func_t func)
836{
837 const unsigned long start_block_id = phys_to_block_id(start);
838 const unsigned long end_block_id = phys_to_block_id(start + size - 1);
839 struct memory_block *mem;
840 unsigned long block_id;
841 int ret = 0;
842
843 if (!size)
844 return 0;
845
846 for (block_id = start_block_id; block_id <= end_block_id; block_id++) {
847 mem = find_memory_block_by_id(block_id);
848 if (!mem)
849 continue;
850
851 ret = func(mem, arg);
852 put_device(&mem->dev);
853 if (ret)
854 break;
855 }
856 return ret;
857}
858
859struct for_each_memory_block_cb_data {
860 walk_memory_blocks_func_t func;
861 void *arg;
862};
863
864static int for_each_memory_block_cb(struct device *dev, void *data)
865{
866 struct memory_block *mem = to_memory_block(dev);
867 struct for_each_memory_block_cb_data *cb_data = data;
868
869 return cb_data->func(mem, cb_data->arg);
870}
871
872/**
873 * for_each_memory_block - walk through all present memory blocks
874 *
875 * @arg: argument passed to func
876 * @func: callback for each memory block walked
877 *
878 * This function walks through all present memory blocks, calling func on
879 * each memory block.
880 *
881 * In case func() returns an error, walking is aborted and the error is
882 * returned.
883 */
884int for_each_memory_block(void *arg, walk_memory_blocks_func_t func)
885{
886 struct for_each_memory_block_cb_data cb_data = {
887 .func = func,
888 .arg = arg,
889 };
890
891 return bus_for_each_dev(&memory_subsys, NULL, &cb_data,
892 for_each_memory_block_cb);
893}